Abstract: Based on chaos theory and nonlinear dynamics,the fault warning diagnosis method is proposed where the maximum Lyapunov exponent is the evaluation index for the operating state,and the Elman-decision tree is used to predict and classify the failures. The method is testified by using a data set proposed by Paderborn University,which is measured from a normal bearing,an outer ring artificially Electrical Discharge Machined（EDM）bearing and outer ring drilled bearing. The result shows that the maximum Lyapunov exponent can effectively distinguish the data sets from different fault states. Then an Elman-decision tree model is established and trained to realize the maximum Lyapunov exponent prediction and fault identification,which is proved of a relatively high accuracy based on the simulation results. The proposed method can achieve good performance even in presence of moderate noise,and can reduce the computation burden by avoiding training multiple corresponding neural networks for all the faults,and it provides a new idea for fault warning diagnosis.